Kotlin / kotlin-spark-api

This projects gives Kotlin bindings and several extensions for Apache Spark. We are looking to have this as a part of Apache Spark 3.x
Apache License 2.0
455 stars 34 forks source link

Question About Mapping #185

Open bkenn opened 1 year ago

bkenn commented 1 year ago

Hello, question for the audience. How should we handle the calculation of fields in a dataset map function? Do we need a manual A -> B class mapping or can we use something generic like the map Row -> Row and then to to handle the field mappings? I can see it becoming tedious to pass in every since property from A -> B. Some of the classes I need to process have 20+ properties.

For example, I would like to accept fields as String?, but in later data frames I want to convert them to Int?. Maybe it would be something like what is below. I know that this example is small but keep in mind I don't want to pass in all 20+ properties to another class constructor. Maybe in the example below we should use sealed classes for possible errors? Not to opinionated about how this should be handled. I know the following code cannot work since RDD's are immutable but it would nice to have some kind of convenience like below to work around that.

data class Client(val age: String?)

data class ClientCalculated(val age: Int?, val errorMessage: String?)

fun litNullAsString() = functions.lit(null).cast(DataTypes.StringType)

 listOf(Client("30"), Client("thirty"))
                .toDS()
                .withColumn("errorMessage", litNullAsString())
                .map {
                    val age = it.getAs<String?>("age")?.toIntOrNull()
                    it["age"] = age
                    it["errorMessage"] = if (age == null) {
                        "age is invalid"
                    } else {
                        null
                    }
                    it
                }.to<ClientCalculated>()
Jolanrensen commented 1 year ago

How would you convert a data class to another data class outside of Spark? I think manually passing the 20+ properties ensures you have the correct names and types all the time. Of course, you could provide some helper function as factory function of the type you want to convert to, which is probably how I would handle it:

data class Client(val extraProp1: Int, val extraProp2: Double, val age: String?)

data class ClientCalculated(val extraProp1: Int, val extraProp2: Double, val age: Int?, val errorMessage: String?) {
    companion object {
        fun fromClient(client: Client): ClientCalculated {
            val age = client.age?.toIntOrNull()
            val errorMessage = if (age == null) "Age is not a number" else null
            return ClientCalculated(client.extraProp1, client.extraProp2, age, errorMessage)
        }
    }
}

withSpark {
    listOf(Client(1, 1.0, "30"), Client(2, 2.0, "thirty"))
        .toDS()
        .map { ClientCalculated.fromClient(it) }
}

Not sure how sealed classes would help here. Also creating a new row with names is unfortunately not possible in spark as far as I know. Curious what other people think too!

asm0dey commented 1 year ago

One more idea: in JetBrains, in our Spark workflows, we usually use raw data frames, and we only require functions to accept and return typed datasets. I just wanted to let you know that you can do the same. This way, in many parts of your workload, you could omit the tedious work :)